Update README.md

README.md

@@ -142,3 +142,72 @@ configs:
   - split: test
     path: humaneval-rkt/test-*
 ---
+
+# Dataset Card for MultiPL-E-fixed (OCaml, Lua, R, Racket, Julia)
+
+This dataset provides corrections for the **OCaml, Lua, R, Racket, and Julia** portions of the [nuprl/MultiPL-E](https://github.com/nuprl/MultiPL-E) benchmark.
+
+### Original Dataset Information
+- **Repository:** https://github.com/nuprl/MultiPL-E
+- **Paper:** https://ieeexplore.ieee.org/abstract/document/10103177
+- **Original Point of Contact:** carolyn.anderson@wellesley.edu, mfeldman@oberlin.edu, a.guha@northeastern.edu
+
+### This Version
+- **Repository:** https://github.com/jsbyun121/MultiPL-E-fixed
+---
+
+## Dataset Summary
+
+MultiPL-E is a large-scale dataset for evaluating code generation models across 22 programming languages. It was created by translating the OpenAI HumanEval and MBPP Python benchmarks into other languages using a compiler-based approach.
+
+However, analysis of the dataset revealed several logical errors, inconsistencies, and language-specific issues in the generated prompts and test cases. These issues can lead to inaccurate evaluation scores by unfairly penalizing models for correctly identifying flaws in the prompts.
+
+This repository provides a **corrected version** of the dataset specifically for **OCaml, Lua, R, Racket, and Julia**. The goal of this version is to provide a more reliable and accurate benchmark for evaluating Large Language Models on these languages.
+
+## Summary of Corrections
+
+The following modifications were made to address issues in the original dataset.
+
+### 1. Logical Problems in Prompts and Test Cases
+Several problems in the HumanEval portion of the dataset were corrected for the following issues:
+
+* **`HumanEval_75_is_multiply_prime`**: Resolved a mismatch between problem instructions and test cases.
+* **`HumanEval_92_any_int`**: Fixed an incorrect test case that did not align with the problem's requirements.
+* **`HumanEval_116_sort_array`**: Corrected a discrepancy between the sorting criteria in the instructions and the test cases.
+* **`HumanEval_128_prod_signs`**: Amended an incorrect example in the prompt's docstring.
+* **`HumanEval_140_fix_spaces`**: Corrected a faulty test case.
+* **`HumanEval_142_sum_squares`**: Repaired corrupted or syntactically incorrect examples.
+* **`HumanEval_145_order_by_points`**: Clarified vague and ambiguous logic in the question to provide a more precise problem statement.
+* **`HumanEval_148_bf`**: Fixed a contradiction between the provided examples and the main instructions.
+* **`HumanEval_151_double_the_difference`**: Replaced an incorrect test case that produced an invalid result.
+* **`HumanEval_162_string_to_md5`**: Addressed unspecified handling for `None`/`null` data types required by the test cases.
+
+### 2. General Prompt Ambiguities
+* **0-Based Indexing:** Added clarifications to prompts where array/list index interpretation was ambiguous, explicitly enforcing a 0-based convention to ensure consistent behavior.
+
+### 3. Language-Specific Fixes
+* **R:** Corrected issues related to the handling of empty vectors, a common edge case.
+* **OCaml:** Fixed incorrect usage of unary operators to align with OCaml's syntax.
+* **Julia:** Resolved parsing issues caused by the triple-quote (`"""`) docstring character.
+* **Lua & Racket:** `[Add a brief, high-level description of the fixes for these languages here.]`
+
+## Using This Dataset
+
+This corrected dataset is designed to be a **drop-in replacement** for the official MultiPL-E data for OCaml, Lua, R, Racket, and Julia.
+
+To use it, simply replace the original `humaneval-[lang]` files with the corrected versions provided in this repository. The data structure remains compatible with standard evaluation frameworks like the [BigCode Code Generation LM Harness](https://github.com/bigcode-project/bigcode-evaluation-harness).
+
+## Citation and Attribution
+
+If you use this corrected version of the dataset in your work, we ask that you please cite the original MultiPL-E paper and also acknowledge this repository for the corrections.
+
+**Original Paper:**
+```bibtex
+@inproceedings{cassano2023multipl,
+  title={MultiPL-E: A Scalable and Extensible Approach to Benchmarking Neural Code Generation},
+  author={Cassano, Federico and Gouwar, John and Nguyen, Daniel and Nguyen, Tuan and Phothilimthana, Phitchaya and Pinckney, David and Anderson, Carolyn and Feldman, Michael and Guha, Arjun},
+  booktitle={2023 IEEE/ACM 20th International Conference on Mining Software Repositories (MSR)},
+  pages={707--719},
+  year={2023},
+  organization={IEEE}
+}